Abstract:Particles characterization is a significant part of numerous studies in material sciences and engineering technologies. Microscopy images of materials containing particles are usually analyzed by operator with manual counting and measuring of particle sizing by a software ruler. Traditional automated image analyzing methods such as edge detection, segmentation, etc. are not universal, giving poor results on noisy pictures and need empirical fitted parameters. To realize automatic method of particles recognition on scanning tunneling microscopy (STM) data we used U-net and modified U-net neural networks, which was trained on ten STM images contained 1918 particles. Verification on 3 pictures with 695 particles showed mAP=0.12 for modified U-net neural network.

Using Computer Vision and Deep Learning for Nanoparticle Recognition on Scanning Probe Microscopy Images: Modified U-net Approach